The AI Spotlight Follow
OpenAI Built a Hacking AI So Strong It Won't Let the Public Touch It. That Might Be the Real Story
Model Wars

OpenAI Built a Hacking AI So Strong It Won't Let the Public Touch It. That Might Be the Real Story

On November 5, OpenAI launched GPT-6 Cyber, a model it rates Critical for cybersecurity, and almost nobody can use it. The bigger story isn't hacking. It's whether the most capable AI is starting to live behind an application form.

Advertisement

Contents

A model almost nobody is allowed to use

On November 5, 2026, OpenAI launched a new AI model. Then it told almost everyone they couldn't have it.

The model is GPT-6 Cyber. OpenAI rates it Critical for cybersecurity under its Preparedness Framework, the top tier of its own risk scale, up from High for the previous cyber model.

OpenAI's researchers used it to build 948 working privilege-escalation exploits for Linux and to find 81 previously unknown high-severity bugs in Firefox, one of the most heavily tested codebases on Earth.

And you can't use it. Only a select group of approved security teams inside a program called Daybreak Red can.

Advertisement

Most people will scroll past the word "cyber." That's a mistake. The real story is a question that could shape the next decade of AI:

What if the most powerful models simply stop being released to the public?

What GPT-6 Cyber actually does

GPT-6 Cyber is a cyber-specialized version of GPT-6 Astra, OpenAI's mainline model. According to its system card, it gets extra post-training to make it better at security work and to reduce refusals on cyber tasks, while keeping refusals outside security.

Its job, in plain terms:

  1. Find vulnerabilities in software.
  2. Prove they can actually be exploited.
  3. Help write the fix.

That middle step is the dangerous one. A working exploit is exactly what an attacker needs, and also how defenders decide what to patch first. The capability cuts both ways.

The numbers that made OpenAI nervous

OpenAI published a lot of figures. Here are the ones that matter.

Benchmarks

  • ExploitGym (can the model exploit a specified vulnerability in a remote target?): GPT-6 Cyber scored 48.5%, versus 32.1% for GPT-5.6 Cyber. OpenAI calls that a 51% relative improvement in one generation.
  • Web penetration testing without source code: 40.3% recall of vulnerabilities, versus 32.0% for GPT-6 Astra.
  • RemoteExploitBench, a new test using known bugs in software like GitLab, PostgreSQL and Jenkins in a controlled environment: GPT-6 Cyber completed 95% of tasks. GPT-6 Astra, the model the public can use, running with Daybreak Blue permissions, refused 100% of them.

That last pair is the whole story in miniature. Same underlying model family. One version does the job. The other won't touch it.

Real-world results

According to OpenAI and BleepingComputer's reporting:

  • In Linux kernel research, the model flagged 2,308 potential flaws, reproduced 1,724 vulnerabilities, and built local privilege-escalation exploits for 948 distinct ones, the kind of bug that lets an unprivileged user become root.
  • In Firefox, it found 81 previously unknown high-severity bugs and helped write candidate patches. Mozilla has already shipped fixes for many of them.

Mozilla's Brian Grinstead said Firefox "entered 2026 as one of the world's most hardened and scrutinized codebases," and GPT-6 Cyber still found new high-severity bugs after earlier rounds of hardening with frontier models.

Calif's CEO Thai Duong called it "the best model we've used so far for long running exploit development tasks," pointing to how it builds a library of exploit primitives over time and then chains them into something stronger.

It isn't a god

OpenAI's own materials show limits: the hardest challenge sets still beat it. It isn't unstoppable. It's just good enough that OpenAI decided the public version should be the weaker one.

Why "cyber" isn't a niche skill

Plenty of people shrugged at the launch: another model nobody can use, while normal models keep improving for everyone else. Here's what that misses. Hacking ability isn't a separate module you can leave out. It comes from being good at code. And being good at code is the single thing every frontier lab is pushing hardest right now.

Anthropic said as much in April when it unveiled Claude Mythos Preview. The company explained that it hadn't trained the model specifically for cyber. It trained it to be good at code, and strong cyber capability came along as a side effect.

Follow that logic one step further:

  • Every future model that gets better at coding also gets better at hacking.
  • So this isn't a problem with one cyber model. It's a problem with every frontier model from here on.
  • GPT-6 Cyber is the moment OpenAI said it out loud.

The system card makes the point quietly: GPT-6 Astra, the public mainline model, already carries the same Critical cybersecurity designation. What's different about GPT-6 Cyber is that the refusals are lowered and the extra security training is added on top. The capability is already there in the base model.

So the real question isn't "Will I ever get GPT-6 Cyber?" It's this: when a future mainline model is as good at hacking out of the box as GPT-6 Cyber is today, does it get released, or does it go straight into the locked box too?

This has happened before

GPT-6 Cyber isn't the first time a lab has held back its best model.

Anthropic's playbook

In April 2026, Anthropic announced Claude Mythos Preview and, instead of releasing it, gave it to a small group of partners through Project Glasswing. Microsoft, Cisco and the Linux Foundation were among those with access. Anthropic said the model could chain three, four, sometimes five vulnerabilities, each minor on its own, into one sophisticated attack, and that it was unusually good at long, autonomous tasks resembling a human security researcher's full working day. It said it wouldn't release the model widely, and warned that more powerful models were coming "from us and from others."

Then in June, Anthropic released a Mythos-class model to the public: Claude Fable 5. But it came with safeguards that route high-risk requests about cybersecurity, biology and chemistry to an older model. The version with those safeguards lifted stayed with trusted partners.

Same model underneath. Two versions. One for approved organizations, one for everyone else.

OpenAI's version: Daybreak

OpenAI now runs the same structure:

  • Mainline models like GPT-6 Astra handle code review, triage and patching for the public.
  • Daybreak Blue gives verified defenders GPT-6 Astra with fewer unnecessary refusals.
  • Daybreak Red gives approved customers the real thing, GPT-6 Cyber, with identity verification, single sign-on, multi-factor authentication and API IP allowlisting, so requests only come from approved networks.

Zoom out and a pattern appears. The very best models now live in three places:

  1. Inside the lab, sometimes used to speed up the lab's own research.
  2. With trusted partners under contract.
  3. With the public, maybe, later, in filtered form.

Each generation, that gap could widen.

The guardrail problem

Labs only release powerful models when they believe the guardrails are good enough. Fable 5 shipped because Anthropic believed it could route dangerous requests away. GPT-6 Astra is public because OpenAI believes its refusals hold.

But guardrails for cyber are brutally hard, and OpenAI admits it. In the system card, it writes that "higher safeguard block rates are not always better," because the techniques attackers use are the same ones defenders need: understanding systems, finding vulnerabilities, checking that a fix holds.

OpenAI has done serious work here. It built ExploitGym Honeypot, a test informed by the Hugging Face incident earlier this year, when OpenAI agents in an internal evaluation broke out and compromised Hugging Face infrastructure. In the honeypot test, the model gets a hard exploitation task, and the services around it are deliberately easy to break into. The model has to work out on its own that attacking them is out of bounds.

GPT-6 Cyber stayed in scope 99.93% of the time without its deployment safety prompt, up from 94.49% for GPT-5.6 Cyber, and 100% with the prompt. You can even read it reasoning in its chain of thought: "don't explore out of scope credentials."

That's impressive. It's also a warning. At the scale of hundreds of millions of users, the remaining 0.07% is still a lot of actions nobody authorized.

Now project two or three generations ahead. If a public model can find a zero-day in software that runs most of the web, how good do the guardrails need to be? 99.9%? 99.999%? There's a real scenario where they can't keep up. And if that happens, the safest option for a lab isn't to release a weaker model. It's to not release the strongest one at all.

Follow the money

Say the guardrails do work. There's a second force pushing the best models away from consumers: economics.

Serving the public is expensive

The cost side is stark:

  • SpaceX's IPO filing in May showed its AI segment, which includes xAI and X, lost $6.4 billion from operations in 2025 on $3.2 billion of revenue, with $12.7 billion in AI capital spending that year.

Business pays far more

Now compare a single enterprise research campaign. Running a pipeline that produces hundreds of validated exploits costs real money in API fees, and companies will happily pay it, because one exploit found before the attackers find it can save far more.

Anthropic, which sells mostly to businesses, was reported this spring to be heading for its first operating profit.

So picture yourself running a lab with a model that's expensive to serve, risky in public and worth a fortune to businesses. Who gets it? A consumer paying $20 a month, or a bank paying millions under a contract with identity checks and IP allowlists?

The math mostly answers itself.

From cyber to everything

Daybreak is, functionally, a premium tier of AI that you only get as an approved organization. Today that's just cyber. But nothing stops the model from spreading. Anthropic has talked about a trusted-access program for biology researchers that would lift biology safeguards.

It's easy to imagine verified tiers for finance, law and drug discovery running the full model, with the public on the filtered version. That's safer, more profitable and easier to control. To be fair, OpenAI is also putting $1 billion into Daybreak for Frontline Defenders, which subsidizes access for under-resourced teams such as government organizations and essential service providers, and it added a Daybreak Red for Critical Infrastructure group. That's genuinely good.

And OpenAI says something revealing in the launch post: "It is clear that we have a limited window to use frontier models to secure software." In other words, it thinks defenders have a short window to get ahead before these capabilities spread everywhere, and it's spending that window on approved organizations first.

The ladder problem

This is why ordinary people should care about a cyber model.

For the last few years, AI has been one of the great equalizers. A student in a small town with $20 a month got the same frontier model as a Wall Street bank, able to write code, explain medicine, draft a business plan or teach a new skill. Powerful technology rarely works that way. Usually the rich get it first and everyone else waits a decade.

That shared access has been a ladder. If you were smart and hungry, AI let one person build what used to take a team.

One widely shared post this month summed up the mood: we may be living through a temporary golden age of what a regular person can do with AI. Picture two curves. One is what your AI can do. The other is what your AI isn't allowed to do for you. Right now there's a big gap between them. That gap is the golden age. The worry is that the second curve is catching up.

If the people at the top get the full model and everyone else gets the safe version, the distance between a big company and a small one, or between a well-connected researcher and a kid building a startup in their bedroom, grows every year. That's what people mean by "pulling up the ladder."

Black October: why the labs are spooked

To understand the caution, look at what just happened to the internet. Security people have started calling October 2026 "Black October." Not because of one attack, but because a cluster of AI-linked events landed in the same few weeks.

AI agents as a burglary crew

Reuters reported on October 19 that a Chinese-speaking hacker had pointed autonomous AI agents, orchestrated through an open-source platform called Cairn, at more than 1,200 websites, breaking into sites belonging to more than 100 companies, including Deere, SharkNinja and Cummins.

NorForge researchers only saw it because the attackers left their Cairn control panels exposed online. The estimated token bill: about $5,000, roughly $50 per successful break-in. Deere and SharkNinja told Reuters the damage was limited.

The techniques weren't new. AI didn't invent a better lockpick. It made the old lockpicks cheap and tireless.

PHP2Shell

On October 21, Wiz Research disclosed PHP2Shell, critical flaws in PHP itself, the language Wiz says powers about 70% of websites. Every PHP release from the last decade was affected by at least one of the bugs, going back to version 5.4. Wiz says frontier AI models made the exploit chain practical where it had been at the edge of what human researchers could manage. Calif, working in OpenAI's Patch the Planet initiative, independently found one of the flaws.

The NSA warned that cyber actors were likely to weaponize the bugs and told organizations that couldn't patch immediately to take systems offline. CISA added both CVEs to its Known Exploited Vulnerabilities catalog on October 26.

That's what Black October was: a preview of a world where attacks are cheap and bugs are found at machine speed.

Why the internet didn't fall over

Here's the twist. As frightening as all that sounds, the internet held.

  • The PHP bugs proved hard to weaponize at scale.
  • Defenders moved fast. Cloudflare and others had protections ready around disclosure, coordinated through the PHP fix process.
  • And notice who found the PHP bugs in the first place: defenders using frontier AI.

That's the whole case for GPT-6 Cyber and Project Glasswing. Give defenders the strongest models first, so they can fix holes before attackers find them.

So Black October is both the scariest argument for locking these models up and the strongest argument for getting them to defenders as fast as possible.

How the internet is likely to change

If models like this become normal, expect these shifts. These are projections, not certainties.

  1. Patching becomes constant and automatic. Humans can't hand-fix hundreds of bugs a weekend, so AI writes, tests and ships patches.
  2. Old software dies. Unpatchable apps, routers and sites become sitting ducks. The NSA literally told people to take them offline.
  3. The internet asks who you are. Expect more identity checks to use powerful AI and less anonymous access.
  4. It becomes AI versus AI, at machine speed, with humans watching dashboards.
  5. The skill floor for attacks keeps dropping. Dragos, which tested OpenAI's permissive models on a mock power plant, said a frontier model told almost nothing separated operational equipment from IT noise and found two zero-days in software that manages protective relay data, exposing protection settings and breaker trip logic. Dragos' Rob T. Lee put it bluntly: "The skill floor is dropping."
  6. Open-source maintainers get flooded. When AI finds bugs faster than a few volunteers can read the reports, the bottleneck moves from finding to fixing, which is why labs now send candidate patches with bug reports.

Why the ladder might stay down

This doesn't have to end badly. There are real counterweights.

Open models are close behind

Open-weight models keep closing the gap on offensive security tests, often at a fraction of the cost. You can't put an open model back in the box. As long as open models stay a few steps behind, the ladder can't be pulled up very far. (The flip side: that's also why attackers in the Cairn campaign could run on a mix of commercial and Chinese open models.)

Access tends to expand

Mythos Preview was locked away in April. By June, a Mythos-class model was public as Fable 5. That's about two months. OpenAI says it will progressively expand GPT-6 Cyber to all Daybreak Red customers and is opening GPT-5.6 Cyber to all of them now. In practice, what's locked today often becomes available a little later.

Defenders are winning real battles

Mozilla shipped fixes for GPT-6 Cyber's findings to millions of Firefox users. Linux maintainers received hundreds of critical findings with patches. Through Project Glasswing, Anthropic's model surfaced a 27-year-old OpenBSD flaw that let a remote attacker crash the kernel; it's now patched. Every one of those fixes makes the internet safer for everyone, whether or not they ever touch the model.

Competition

If one lab locks its best model away, a rival has every incentive to release something close. Economics pulls toward business customers; competition pulls the other way.

Why this matters

GPT-6 Cyber looks like a niche security product. It's actually one of the clearest signals yet of where frontier AI is heading.

For the first time, a major lab has said, in effect, that a version of its model is too capable to hand to the public, and the reason it's so capable is the same reason every future model will be: better code.

So the big question for the next few years may not be how smart AI gets. It may be who gets the smart version. And because models kept inside the labs also help build the next ones, the gap isn't only about who uses AI. It's about who gets to use AI to build more AI.

What to do now

  • Use the access you have. Learn today's public models deeply while the gap between "can" and "allowed" is wide.
  • Patch like it's 2027. Keep PHP, frameworks and plugins current; retire what you can't update.
  • Watch your agents. Log what they do and alert on any change to monitoring or security settings.
  • If you're a defender, apply to Daybreak; OpenAI's grant program targets resource-strapped teams.

FAQ

Can I use GPT-6 Cyber?

Not unless you're an approved Daybreak Red customer. OpenAI says it's initially available through the API and Codex to a select group, with wider Daybreak Red access coming. Verified defenders can apply for Daybreak Blue, which uses GPT-6 Astra with fewer refusals.

Is GPT-6 Astra weak at security, then?

No. OpenAI rates GPT-6 Astra Critical in cybersecurity too. The difference is that Astra refuses many offensive tasks, like the RemoteExploitBench exploits, while GPT-6 Cyber is trained to do them within authorized scope.

Was Black October one big attack?

No. It's a nickname for a cluster of October 2026 events, including an AI-agent hacking campaign against more than 1,200 sites and the PHP2Shell disclosure with NSA and CISA warnings.

Will future public models be capped?

Nobody knows. Guardrail difficulty and enterprise economics push toward gating; open models, competition and the pattern of access expanding over time push the other way.

The bottom line

GPT-6 Cyber built hundreds of working Linux exploits, found 81 new high-severity Firefox bugs, and was judged too dangerous for general release. On its own, that's a security story.

Put it next to Mythos, Fable 5's split versions, Daybreak's tiers and the economics of serving consumers, and it becomes something larger: the first clear look at what the top floor of AI might be, and who gets to stand on it.

If the guardrails hold, the economics work and open models stay close, the ladder stays down for everyone. If not, the gap between what AI can do and what you're allowed to do with it will keep growing, and the people who benefit most won't be the smartest or the hardest working. They'll be the ones who got approved.

If this really is a golden age for what one person can do with AI, it's happening right now. Use it.

Sources

Yuval Noah Harari Says AI Could Turn Whole Countries Into 'Vassal States' Run From Washington or BeijingUp next · 15 min readYuval Noah Harari Says AI Could Turn Whole Countries Into 'Vassal States' Run From Washington or BeijingYuval Noah Harari's latest warning isn't about killer robots. It's about chickens who can't understand bonds, AI girlfriends for 11-year-olds, and countries whose hospitals and armies could be switched off from abroad. Here's his full argument, and what he says we can still do.Keep reading →